Project Meeting 2021.04.27

Technical Meeting

• issues #419, #377 updated
• release complete, activitysim version 0.9.9.1 and populationsim version 0.5
• chunking re-factoring going well
  • added a monitoring thread that runs in parallel and checks RSS memory usage every sec
  • currently running 100k HHs with MTC example on Jeff's laptop
  • next try full scale examples on our Windows servers and add support for multizone models
  • also has settings reload feature Alex suggested
  • OS memory report versus numpy/pandas are typically +/- 20% of one another
  • When numpy/pandas decides to finally make a copy of the data in memory within an operation or move some hunk of memory around in the heap are the unpredictable moments that make this exercise difficult
  • so dampening the sensor so we don't get too much yo-yo-ing
  • We'll want to update all the examples and documentation to drop chunking and switch to this assuming it works well
• MTC talking to Clipper folks about any subsidy data that might be available
• sandag testing
  • separate support work is complete
  • sandag is working on trying one zone and two zone setups as well
  • performance of one zone, two zone not much better?
  • would be good to start from psrc two zone example since its more complete
  • share the runtimes by submodel and then let's think about spots that we could make faster
  • semcog will try a 1 zone and 3 zone setup and can share their experience
  • can we do presampling for one zone system? its only for two and three zone system models because they have mazs
  • presampling does taz level sampling and then picks an maz for each taz sampled so its very similar to the one zone approach